These two threads go into some detail with the problems with similar models.
In general, the problem is that the parameters are not very strongly identified. For instance, within the exponent, every increase in D can be offset by a decrease in A. That means that if you don’t have enough data (in a subgroup) to estimate the plateau, A, D is not identified either. One option could be to only allow for hierarchical parameters on D or A but not for both if you can justify it. Another option is to reparametrize in terms of A and ratio = D/A. If A > 0, then the maximum derivative scaled by the maximum value could make sense.